def __init__(self, search_space):
search_config = {}
search_config['learning_rate'] = ag.Real(1e-3, 1e-2, log=True)
search_config['epochs'] = ag.Choice(40, 80)
for config in search_config.keys():
if not config in search_space.keys():
search_space[config] = search_config[config]
self.search_space = search_space
def __init__(self,dictionary_of_hyperparameters):
search_config = {}
search_config['learning_rate'] = ag.Real(1e-3, 1e-2, log=True)
search_config['epochs'] = ag.Choice(40, 80)
self.dictionary_of_hyperparameters = dictionary_of_hyperparameters
for config in search_config.keys():
if not config in self.dictionary_of_hyperparameters .keys():
self.dictionary_of_hyperparameters [config] = search_config[config]
self.dataset = dataset
self.init_estimator = SimpleFeedForwardEstimator(
num_hidden_dimensions=[10],
prediction_length=dataset.metadata.prediction_length,
context_length=100,
freq=dataset.metadata.freq,
trainer=Trainer(ctx="cpu",
epochs=5,
learning_rate=1e-3,
num_batches_per_epoch=100
)
)
transformation = estimator.create_transformation()
dtype = np.float32
num_workers = None
num_prefetch = None
shuffle_buffer_length = None
trainer = Trainer(ctx="cpu",
epochs=1,
learning_rate=0.01,
num_batches_per_epoch=100
)
self.training_data_loader = TrainDataLoader(
dataset=dataset.train,
transform=transformation,
batch_size=trainer.batch_size,
num_batches_per_epoch=trainer.num_batches_per_epoch,
ctx=trainer.ctx,
dtype=dtype,
num_workers=num_workers,
num_prefetch=num_prefetch,
)
"""Train a BERT model using AutoGluon."""
import sys
import autogluon as ag
from bert_model import train_bert
@ag.args(
data_file=sys.argv[1],
epochs=ag.Choice(40, 80),
lr=ag.Real(1e-5, 1e-4, log=True),
batch_size=ag.Choice(256, 512, 1024),
wd=ag.Real(1e-3, 10, log=True),
num_heads=ag.Choice(8, 16, 32, 64),
num_pred_hiddens=ag.Choice(256, 512, 1024, 2048),
ffn_num_hiddens=ag.Choice(512, 1024, 2048),
num_layers=ag.Choice(6, 8, 12),
dropout=ag.Real(1e-1, 8e-1, log=True),
)
def run_training(args, reporter):
"""Launch training process."""
args.num_hiddens = args.num_heads
return train_bert(args, reporter)
def run():
"""Run tuning."""
scheduler = ag.scheduler.FIFOScheduler(run_training,
resource={
'num_cpus': 1,
from asset import optimizer
from gluonts.dataset.loader import TrainDataLoader
from gluonts.trainer import Trainer
with tqdm(training_data_loader) as it:
for batch_no, data_entry in enumerate(it, start=1):
if False:
break
inputs = [data_entry[k] for k in input_names]
dictionary_of_hyperparameters = {}
#dictionary_of_hyperparameters ['learning_rate'] = ag.Real(1e-3, 1e-2, log=True)
dictionary_of_hyperparameters['epochs']=ag.Choice(10, 20)
class auto:
def __init__(self,dictionary_of_hyperparameters):
search_config = {}
search_config['learning_rate'] = ag.Real(1e-3, 1e-2, log=True)
search_config['epochs'] = ag.Choice(40, 80)
self.dictionary_of_hyperparameters = dictionary_of_hyperparameters
for config in search_config.keys():
if not config in self.dictionary_of_hyperparameters .keys():
self.dictionary_of_hyperparameters [config] = search_config[config]
self.dataset = dataset
self.init_estimator = SimpleFeedForwardEstimator(
num_hidden_dimensions=[10],
prediction_length=dataset.metadata.prediction_length,